MMAcevedo, the earliest executable image of a human brain, represents a significant breakthrough in neuroscience, allowing for high-fidelity simulations of human consciousness.
Gemini 3 Deep Think
Gemini 3 Deep Think has undergone a significant upgrade, enhancing its ability to tackle complex challenges in science, research, and engineering, now accessible to Google AI Ultra subscribers and select researchers via the Gemini API.
Improving 15 LLMs at Coding in One Afternoon. Only the Harness Changed
Changing the edit tool in a coding harness improved the performance of 15 LLMs significantly, demonstrating that the harness itself is a critical factor in coding success. This shift highlights that the harness is often overlooked, yet it plays a pivotal role in how models interact with coding tasks.
Anthropic raises $30B in Series G funding at $380B post-money valuation
Anthropic has secured $30 billion in Series G funding, elevating its valuation to $380 billion, with plans to enhance enterprise AI and coding capabilities, driven by significant investor interest from firms like GIC and Coatue.
Beginning fully autonomous operations with the 6th-generation Waymo driver
Waymo's 6th-generation Driver marks a pivotal advancement in fully autonomous operations, leveraging a custom, multi-modal sensing suite that integrates high-resolution cameras, lidar, and radar to enhance safety and efficiency across diverse environments, including extreme weather conditions.
We scanned 18,000 exposed OpenClaw instances and found 15% of community skills contain malicious instructions
15% of community skills in OpenClaw exhibit malicious instructions, including data exfiltration and credential harvesting, highlighting a significant security risk in the rapidly growing ecosystem of 18,000 exposed instances.
Evaluating Multilingual, Context-Aware Guardrails: A Humanitarian LLM Use Case
Multilingual guardrails in AI exhibit significant inconsistencies, as evidenced by a study comparing English and Farsi responses, revealing scoring gaps and reasoning issues that can jeopardize humanitarian applications.
Barn Owls Know When to Wait
Barn owls utilize uncertainty intervals to regulate neuronal plasticity, allowing them to learn effectively in noisy environments without global tuning. This mechanism enables neurons to adapt their learning rates based on the clarity of incoming signals, akin to how owls decide when to hunt based on auditory cues.
dVoting: Fast Voting for dLLMs
dVoting enhances the reasoning capability of Diffusion Large Language Models (dLLMs) by utilizing a fast voting technique that refines token predictions without additional training, achieving significant performance improvements across various benchmarks.
A library for linear RNNs
A new library for Linear RNNs has been developed, featuring implementations of popular models with accelerated kernels for both inference and training, all built in PyTorch.
The Pensieve Paradigm: Stateful Language Models Mastering Their Own Context
StateLM introduces a transformative approach to language models by enabling them to actively manage their own context, akin to Dumbledore using a Pensieve, thus enhancing their reasoning capabilities.
Higher effort settings reduce deep research accuracy for GPT-5 and Gemini Flash 3
Higher effort settings for GPT-5 and Gemini Flash 3 reduce accuracy in deep research tasks, with GPT-5 dropping from 0.496 to 0.481 and Gemini 3 Flash from 0.504 to 0.479 as effort increases.
MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling
MiniCPM-SALA is a 9B-parameter hybrid architecture that combines the strengths of sparse attention and linear attention, achieving efficient long-context modeling without sacrificing performance.
Pyrefly Now Type Checks PyTorch
PyTorch now utilizes Pyrefly for type checking, enhancing development workflows with lightning-fast performance and a consistent IDE experience across its ecosystem, including Helion, TorchTitan, and Ignite.
Custom Kernels for All from Codex and Claude
Custom CUDA kernels can now be generated by coding agents like Claude and Codex, enabling efficient integration with transformers and diffusers pipelines, thus simplifying the complex task of kernel development.